Deep Learning based Novel Anomaly Detection Methods for Diabetic Retinopathy Screening

Abstract

Computer-Aided Screening (CAS) systems are getting popularity in disease diagnosis. Modern CAS systems exploit data driven machine learning algorithms including supervised and unsupervised methods. In medical imaging, annotating pathological samples are much harder and time consuming work than healthy samples. Therefore, there is always an abundance of healthy samples and scarcity of annotated and labelled pathological samples. Unsupervised anomaly detection algorithms can be implemented for the development of CAS system using the largely available healthy samples, especially when disease/nodisease decision is important for screening. This thesis proposes unsupervised machine learning methodologies for anomaly detection in retinal fundus images. A novel patchbased image reconstructor architecture for DR detection is presented, that addresses the shortcomings of standard autoencoders-based reconstructors. Furthermore, a full-size image based anomaly map generation methodology is presented, where the potential DR lesions can be visualized at the pixel-level. Afterwards, a novel methodology is proposed to extend the patch-based architecture to a fully-convolutional architecture for one-shot full-size image reconstruction. Finally, a novel methodology for supervised DR classification is proposed that utilizes the anomaly maps.